What a learning module's hypotheses were worth, in the terms a paper uses.

Mirrors frameworks/models/mixins/no_reset_evidence.py::TheoreticalLimitLMLoggingMixin, the mixin Monty's unsupervised-inference learning module is built from. It adds four things to a step's log, none of which the module needs in order to recognise anything — they exist so an episode can be judged as well as read:

• "max_evidence" — the strongest hypothesis of every object;
• "target_object_theoretical_limit" — the pose error of the best hypothesis of the target object, i.e. how right Monty could have been if it had picked the best one instead of the most likely one;
• "target_object_pose_error" — the pose error of the most likely hypothesis, which is the error the episode is actually charged with;
• "hypotheses_updater_telemetry" — per object, the whole hypothesis space (Monty.Frameworks.Models.Mixins.NoResetEvidence.HypothesesUpdaterGraphTelemetry).

The gap between the two pose errors is the finding

The theoretical limit is an oracle's score and the MLH's error is Monty's. When they diverge, the hypotheses are fine and the choice among them is not — which is a different repair from "the model never had a good explanation". That is why the mixin computes both rather than one.

Differences from Monty, on purpose

• No inheritance, so no __init_subclass__ guard. Monty refuses at class definition time to mix this into anything but an EvidenceGraphLM; here the guard is the argument's type, and add_detailed_stats/2 has no clause for anything else.
• Monty's mixin replaces the learning module's own detailed statistics rather than adding to them: its _add_detailed_stats does not call super(), so possible_locations, evidences, possible_rotations and symmetry_evidence are absent from the log of the learning module that mixes it in. That is Monty's behaviour, kept.